---
id: "2026-11-02-western-startups-decensor-chinese-models"
url: "https://postcutoff.com/e/2026-11-02-western-startups-decensor-chinese-models/"
as_of: "2026-11-09T21:13:00+01:00"
date: "2026-11-02"
date_precision: day
category: open-source
importance: 2
confidence: high
status: [Confirmed]
sources: 1
editor: human editor
human_review: null
version: "2026-11-04"
---

As of: 2026-11-09 21:13 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-11-02-western-startups-decensor-chinese-models/

# Rest of World: Western startups strip censorship from Chinese open models for enterprise customers

Rest of World reported on Nov 2, 2026 that firms such as Israel's Hirundo and Boston's Lazarus AI use unlearning and weight edits to remove state-aligned censorship from Chinese open models like Qwen; Lazarus's 'ReAligned' models run at about a dozen US companies.

## Key facts
- Hirundo found Qwen avoids Tiananmen, Hong Kong 2019 protests and Xinjiang camps
- Lazarus used topics from leaked censorship guidelines, with answers generated by Gemini Flash
- Airbnb, Uber and Harvey have built models on Chinese bases; OpenEuroLLM is built on Qwen

## What happened

By Rina Chandran and Viola Zhou. Sceptics quoted: Oriane Peter (King's College London) calls universal safety "an impossible standard";
Graham Webster (Stanford DigiChina) says censorship matters most in sensitive regions.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 186 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 125 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 216 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 155 days after its cutoff

## Sources

1. [Rest of World: Startups recalibrate Chinese AI for the West](https://restofworld.org/2026/western-startups-realigning-chinese-ai-models/) (restofworld.org, press)

## Changes

- 2026-11-04 (filed): Created

## Related

- 2026-11-03: [Revolut says self-hosted GLM models cut its AI coding inference costs about 90%, replacing Claude Code and Codex as default](https://postcutoff.com/e/2026-11-03-revolut-self-hosted-glm-cuts-coding-costs/index.md)